Epigenetic profiling for the molecular classification of metastatic brain tumors

Abstract

Optimal treatment of brain metastases is often hindered by limitations in diagnostic capabilities. To meet this challenge, here we profile DNA methylomes of the three most frequent types of brain metastases: melanoma, breast, and lung cancers (n = 96). Using supervised machine learning and integration of DNA methylomes from normal, primary, and metastatic tumor specimens (n = 1860), we unravel epigenetic signatures specific to each type of metastatic brain tumor and constructed a three-step DNA methylation-based classifier (BrainMETH) that categorizes brain metastases according to the tissue of origin and therapeutically relevant subtypes. BrainMETH predictions are supported by routine histopathologic evaluation. We further characterize and validate the most predictive genomic regions in a large cohort of brain tumors (n = 165) using quantitative-methylation-specific PCR. Our study highlights the importance of brain tumor-defining epigenetic alterations, which can be utilized to further develop DNA methylation profiling as a critical tool in the histomolecular stratification of patients with brain metastases.

Fig. 1

Genome-wide DNA methylation profiling of brain tumors. a Diagram describing the normalization algorithm of Infinium HM450K probes. b Matrix depicting the Spearman’s ρ correlation coefficients among primary and metastatic brain tumors. c Principal component analysis (PCA) of GBM (n = 60) and brain metastases (n = 94) using DNA methylation level of 22,483 randomly-selected genomic regions. d DNA methylation level of the 12 most differentially methylated regions between GBM and BM specimens. The upper panel shows the β-value of six genomic regions differentially hypermethylated in BM (cg07076109, cg19111287, cg09962377, cg10982851, cg15002250, and cg23108580) and the lower panel shows the β-value of six genomic regions differentially hypermethylated in GBM (cg25814383, cg26306329, cg06663644, cg20604286, cg04314308, and cg26306994) for each specimen in the study (n = 154). The top and bottom of each box represent the first and third quartile, respectively; the internal line represents the median. e Validation of differentially methylated regions by qMSP in an independent cohort of brain tumor specimens (n = 72). The left boxplot shows the methylation level of a genomic region hypermethylated in BM specimens (BM-C; cg09962377) and the right boxplot the level of a genomic region hypermethylated in GBM specimens (GBM-A; cg25814383; chr19:19,336,240). The top and bottom of each box represent the first and third quartile, respectively; the internal line represents the median. *Wilcoxon test, P-value < 0.02. f ROC showing the prediction potential of brain tumor type using qMSP scores for each independent genomic region and score for the combination (DNAm level of BM-C minus DNAm level of GBM-A; AUC = 0.92, 95% CI = 0.85-0.99); see Supplementary Data 3 for details about validated genomic regions. The AUC values are indicated between square brackets

Fig. 2

DNA methylation differences among brain metastases. a Boxplot representing overall Spearman’s ρ correlation coefficients among brain tumors. The left plot includes all the GBM specimens (n = 60, mean ρ = 0.90 ± 0.06) and all the BM specimens (n = 94, mean ρ = 0.77 ± 0.02). The right plot describes the overall Spearman’s ρ correlation coefficient among BM specimens with anatomical pathology confirmed tumor of origin (BCBM n = 28, LCBM n = 18, and MBM n = 44). The top and bottom of each box represent the first and third quartile, respectively; the internal line represents the median. ***Spearman’s correlation test; P-value < 0.001. b Unsupervised hierarchical clustering using Euclidean distance of the top 5000 most variable genomic regions. c PCA using 31,818 CpG sites with significant (ANOVA, Bonferroni adjusted P-value < 0.05; Supplementary Data 4) differential DNA methylation level among BM with anatomical pathology confirmed tissue of origin (n = 90). d PCA using the differentially methylated region including four BM specimens with uncertain primary tumor of origin (n = 94). e PCA including BM specimens from female patients (n = 58)

Fig. 3

DNA methylation classifiers to predict the tissue of origin of brain metastases. a CV performance across 100 repeats for the RF classifiers to predict tumor of origin, in order of decreasing number of features used in model construction, from left to right (x-axis). Red bars on the boxplots indicate medians and light blue bars depict the performance based on permuted class labels and represent the random background distribution. b Bar plots depicting the prediction performance as measured by sensitivity and specificity for each of the BM types. The bars show the average performance and interquartile range (error bars) across all models with 40 features or more across all repeats. c Bar plots depicting the RF feature importance (mean decreases in Gini impurity score; GIS) of the 15 most predictive genomic regions averaged across all models with 40 features or more and across all repeats. d For three genomic regions in the top 15: boxplots of DNAm β-values across our cohort stratified by tumor of origin (BCBM n = 28, LCBM n = 22, and MBM n = 44) in the upper panels and TCGA cohorts of primary breast tumors (n = 401), primary lung tumors (n = 307), and primary melanomas (n = 83) in the lower panels. Differences in the DNAm levels among the groups were statistically significant for all the cases (Kruskal–Wallis test; P-value < 0.0001). e DNAm levels assessed by qMSP for three regions differentially methylated among the three BM types (n = 59). The top and bottom of each box represent the first and third quartile, respectively; the internal line represents the median. ***Wilcoxon test; P-value < 0.001. f ROC curves showing the prediction potential for the tumor of origin (n = 59) for each of the differentially methylated regions and combinations into BM type-specific scores: MBMscore = DNAm level of MBM-B minus DNAm level of LCBM-C minus DNAm level of BCBM-C; LCBMscore = DNAm level of LCBM-C minus DNAm level of BCBM-C minus DNAm level of MBM-B; and BCBMscore = DNAm level of BCBM-C minus DNAm level of LCBM-C minus DNAm level of MBM-B; see Supplementary Data 6 for details about these genomic regions. The AUC values are indicated between square brackets

Fig. 4

DNA methylation differences among BCBM subtypes. a Hierarchical cluster analysis using Euclidean distance for the DNAm level of 409 regions significantly differentially methylated (one-way ANOVA; FDR-corrected q-value < 0.0005; Supplementary Data 7) among the three breast cancer molecular subtypes (n = 24 BCBM specimens). This analysis revealed three distinct clusters of genomic regions. CL1 includes regions specifically methylated in HR+/HER2− BCBM, CL2 includes regions methylated in both, HR+/HER2− BCBM and HER2+ BCBM, and CL3 includes regions specifically methylated in HR−/HER2− BCBM. b Two-dimensional projection depicting the DCA for differentially methylated regions (n = 409) and BCBM specimens (n = 24) with known IHC profile. This plot shows the spatial overlapping of genomic regions with relative importance for each BCBM molecular subtype. c PCA including 24 BCBM specimens with known molecular subtypes and four BCBM with missing IHC information using 126 genomic regions with classification potential. The unconfirmed specimens were assigned to two different clusters. BCBM-10 and BCBM-23 overlapped with HR+/HER2− BCBM and BCBM-20 and BCBM-31 overlapped with HER2+ BCBM. d IHC evaluation in a CLIA-certified Pathology Department for ER, PgR, and HER2 expression (scale bar, 100 µm). The results confirm the DNAm-based prediction for the expression of HR and HER2. e Magnetic resonance imaging showing two patients with synchronous (case 1) BCBM lesions (BCBM-03 and BCBM-04) and asynchronous (case 2) BCBM lesions (BCBM-05 and BCBM-19). f Phylogenetic tree generated using the Euclidian metric distance for BCBM according to DNAm profile of the 126 genomic regions

Fig. 5

A DNA methylation-based classifier to predict BCBM subtypes. a Boxplots describing the CV performance across 100 repeats for RF classifiers to predict breast cancer subtypes. From left to right, decreasing numbers of features were used to construct the model (x-axis). The top and bottom of each box represent the first and third quartile, respectively; the internal red line represents the median values. Light blue bars depict the performance based on permuted class labels and represent the random background distribution. b Bar plots depicting the prediction performance as measured by sensitivity and specificity for each of the three subtypes. The bars show the average performance and interquartile range (error bars) across all models with 10 features or more across all repeats (HR+/HER2−; n = 12, HER2+; n = 11, and HR−/HER2−; n = 5). c Bar plots depicting the RF feature importance (mean decreases in Gini impurity score; GIS) of the 15 most predictive regions averaged across all models with 10 or more features and across all repeats. d Boxplots of DNAm levels (β-values) across our cohort stratified by subtypes (HR+/HER2−; n = 12, HER2+; n = 11, and HR−/HER2−; n = 5) and TCGA-BRCA cohort of primary breast tumors stratified to match our molecular subtype definitions (HR+/HER2−; n = 443, HER2+; n = 83, and HR−/HER2−; n = 117). Differences in the DNAm levels among the groups were statistically significant for all the cases (Kruskal–Wallis test; P-value < 0.0001). e DNAm levels assessed by qMSP for six genomic regions with differential DNAm among the three BCBM subtypes (n = 31). ***Wilcoxon test; P-value < 0.001. f ROC curves showing the prediction potential for the breast cancer subtypes (n = 31) for each of the six differentially methylated regions and combinations of three or six regions into BCBM molecular subtype-specific scores: HR+/HER2-score = DNAm level of HR+/HER2− minus DNAm level of HER2+ minus DNAm level of HR−/HER2; HER2+ score = DNAm level of HER2+ minus DNAm level of HR+/HER2− minus DNAm level of HR−/HER2; and HR−/HER2-score = DNAm level of HR−/HER2− minus DNAm level of HR+/HER2− minus DNAm level of HER2+; see Supplementary Data 10 for details about these regions. The AUC values are indicated between square brackets

Fig. 6

Summary of the BrainMETH classifiers. a–c BrainMETH classifiers designed to discriminate between primary and metastatic brain tumors (Classifier A), among BM from different tumor of origin (Classifier B), and among BCBM from different molecular subtypes (Classifier C). A set of relevant Illumina probes, genomic regions, and validated primer sets by qMSP is provided for each step of the BrainMETH classifier